基于改进HOG的视频中运动人体特征检测与行为识别

发布时间：2018-12-16 20:14

【摘要】：在当今的计算机视觉领域,动态视频中人体目标的检测与人体正在进行的行为的识别已经成为了一个重要的研究方向。其在智能监控、视频检索、人体运动分析方面有着重要的应用。但是,由于人体的非刚性以及视频背景的复杂性,该方向仍然是一个具有挑战的研究方向。视频中运动人体行为识别过程分为三个过程,即人体目标检测、人体目标特征提取、人体行为识别。梯度方向直方图(HOG)算法诞生之初是用于人体目标检测,并非用于人体行为识别之中,本文将HOG算法与其他算法结合应用于人体行为识别当中。本文通过提出改进的混合高斯背景模型与改进的HOG算法来对人体行为进行识别,主要对跑、走、弯腰、上跳、单手挥动、双手挥动、双脚跳七个人体行为的基本动作进行识别,并在识别上取得了较好的识别效果。本文所做的主要工作如下:在人体目标检测中,首先阐述了几种基本的图像预处理方法,只有将输入的图像进行预处理后才能得到更好的输入图像。因此在此方面介绍了图像增强、图像去噪、数学形态学在图像中的应用。通过图像去噪可以去除图像中影响结果的噪声,通过数学形态学处理,消除了图像中的孤立像素点和人体目标的空洞。介绍了几种常用的背景建模模型、通过对比它们的优缺点,使用了混合高斯模型,但是该模型虽然能够有效的获取到视频中的背景,但是该方法受图像中的阴影影响较大,因此本文提出了一种改进的混合高斯模型即基于HSV空间与阴影抑制方法的混合高斯背景模型,通过将原图像转换为HSV空间中的图像,并通过V向量上的值来有效的抑制阴影。在人体特征提取方面,首先提出了重心特征,通过对人体轮廓的提取,并对人体轮廓的重心计算,得到了重心特征向量。通过对两种直方图的学习,包括梯度方向直方图(HOG)与光流场直方图(HOF),得出单一的某个直方图并不能更好的表现出图像人体的运动特征,因此本文提出了一种改进的HOG算法,即基于HOG、HOF、重心特征的人体特征提取方法。在人体行为识别方面,本文介绍了几种常用的人体行为分类方法,即模版匹配法、直接分类法和状态空间法,通过对几种算法的学习,本文最终选用直接分类法中的KNN算法,并实现了基于HOG、HOF、重心特征的人体特征提取方法的人体行为识别,通过三种特征不同的组合,对Weizmann数据库中7种不同的人体行为动作分别进行了6组不同的实验,证明了本文提出的改进的HOG算法的正确性,识别率达到95.42%。
[Abstract]:In the field of computer vision, the detection of human body and the recognition of human behavior in dynamic video have become an important research direction. It has important applications in intelligent surveillance, video retrieval and human motion analysis. However, due to the nonrigid human body and the complexity of video background, this direction is still a challenging research direction. The recognition process of moving human behavior in video is divided into three processes: human target detection, human target feature extraction, and human behavior recognition. The gradient direction histogram (HOG) algorithm was first used for human object detection, not for human behavior recognition. In this paper, the HOG algorithm is combined with other algorithms in human behavior recognition. In this paper, we propose an improved mixed Gao Si background model and an improved HOG algorithm to identify human behavior, which mainly recognizes the basic actions of running, walking, bending, jumping, waving with one hand and jumping with both feet. Good recognition effect has been achieved in recognition. The main work of this paper is as follows: in human object detection, several basic image preprocessing methods are introduced, only after the input image is preprocessed can the better input image be obtained. Therefore, the application of image enhancement, image denoising and mathematical morphology in image is introduced. Image denoising can remove the noise that affects the result of the image and eliminate the voids of isolated pixels and human body targets by mathematical morphology processing. This paper introduces several common background modeling models. By comparing their advantages and disadvantages, the mixed Gao Si model is used. Although the model can get the background of the video effectively, the method is greatly affected by the shadow in the image. In this paper, an improved mixed Gao Si model is proposed, which is based on HSV space and shadow suppression method. The original image is converted into the image in HSV space, and the shadow is effectively suppressed by the value of V vector. In the aspect of human body feature extraction, the barycenter feature is first put forward, and the barycenter feature vector is obtained by extracting the human body contour and calculating the center of gravity of the human body contour. By studying two histograms, including gradient histogram (HOG) and optical flow field histogram (HOF), it is concluded that a single histogram can not better show the motion characteristics of the image body. Therefore, an improved HOG algorithm based on HOG,HOF, barycenter feature is proposed in this paper. In the aspect of human behavior recognition, this paper introduces several common classification methods of human behavior, that is, template matching, direct classification and state space. Through the learning of several algorithms, the KNN algorithm of direct classification is selected in this paper. The human behavior recognition based on HOG,HOF, barycenter feature extraction method is realized. Through the combination of three different features, six different experiments are carried out on seven different human behavior actions in Weizmann database. It is proved that the improved HOG algorithm is correct and the recognition rate is 95.42.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】